What actually goes wrong in multi-agent teams?
The most comprehensive study available on this question analyzed over 1,600 annotated failure traces from 7 popular multi-agent frameworks and found that the dominant failure category is system design issues — which includes poor planning, unclear role definitions, and flawed task decomposition [3]. These accounted for the largest share of failures, followed by inter-agent misalignment (agents working at cross-purposes). Task verification errors (checking whether work is correct) were a smaller but still significant category. Tool-use errors were not identified as a major failure mode in this taxonomy.
A separate study that built a dataset of 9,533 error-injected agent trajectories to train error-detection models found that planning failures are both the most common and the most difficult to attribute to a specific agent [2]. The researchers had to design special methods to detect these failures because they often look like normal behavior until the consequences compound. This reinforces the finding that planning — not tool use — is the core vulnerability.
Why does planning fail so often?
The failures are not random. One analysis argues that multi-agent systems are not failing because of limited intelligence but because they lack governance over how agents interact [4]. As interaction speed, coupling depth, and role dependency increase without explicit constraints, instability emerges predictably. This paper identifies specific failure patterns like 'epistemic echo loops' (agents reinforcing each other's wrong assumptions) and 'role drift' (agents straying from their assigned function) — both are planning/coordination failures, not tool errors.
Even in constrained physical domains like warehouse robotics, planning failures are the critical bottleneck. A study on lifelong multi-agent path finding found that when planning algorithms fail to produce a solution within time limits, the system must have explicit 'fail policies' to avoid collisions and throughput collapse [5]. The researchers note that while industry practitioners already use such fallback policies, this area had not been formally studied — suggesting that planning failures are a known but under-researched problem.
Do tool-use errors matter at all?
Tool-use errors are real but secondary. The large-scale failure taxonomy study [3] did not list tool-use as a separate failure mode — it was subsumed under system design or task verification. The error-injection study [2] found that tool-related errors (e.g., malformed API calls, wrong parameter passing) are easier to detect and attribute than planning errors, which is why they cause less overall damage. A practical demonstration from the Claw AI Lab platform showed that connecting agents to real codebases and datasets (the 'Claw-Code Harness') reduced common failure modes like partial runs and malformed result reporting [1] — but these are execution issues, not the root cause of team failure.
The evidence across all five studies converges: planning and coordination are the primary failure modes. Tool-use errors are more visible but less consequential. The practical implication is clear: if you are building or debugging a multi-agent system, invest in role definitions, task decomposition, and inter-agent communication protocols before worrying about tool access.
About These Sources
This answer is built on 5 studies (1 peer-reviewed, 4 preprints) — published from 2023 to 2026, 4 from 2024 or later — selected as the most relevant from 5 studies that passed quality screening, drawn from 32 papers retrieved from a database of over 500 million.
Sources used in this answer
Claw AI Lab: An Autonomous Multi-Agent Research Team
Claw AI Lab's platform evaluation on 5 AI research case studies found that connecting agents to real codebases and datasets via the Claw-Code Harness reduced common failure modes like partial runs and malformed result reporting, but the study focused on execution integration rather than root-cause failure analysis [1].
Aegis: Automated Error Generation and Attribution for Multi-Agent Systems
Aegis constructed a dataset of 9,533 error-injected agent trajectories and found that planning failures are the most common and hardest to attribute to a specific agent, requiring specialized detection methods [2].
Why Do Multi-Agent LLM Systems Fail?
Analysis of 1,600+ failure traces across 7 multi-agent frameworks identified 14 failure modes clustered into 3 categories, with system design issues (including planning) as the largest category, followed by inter-agent misalignment, and task verification [3].
Multi-Agent Systems Are Not Failing — They Are Unregulated
Argues that multi-agent instability is a structural governance problem, not a capability problem, identifying specific failure patterns like epistemic echo loops and role drift that emerge from unregulated interaction topologies [4].
Adapting to Planning Failures in Lifelong Multi-Agent Path Finding
In lifelong multi-agent path finding, planning failures that exceed time limits are a critical bottleneck requiring explicit fail policies to avoid collisions and throughput collapse; the study notes this area had not been formally researched despite being used in industry [5].